Traffic lights detection based on enhanced YOLOv5

This study introduces a refined YOLOv5 model aimed at overcoming challenges associated with the detection of diminutive traffic signal lights in intricate environmental settings. The Normalized Wasserstein Distance (NWD) metric is employed to counteract the Intersection over Union (IoU) metric's susceptibility to positional discrepancies in the detection of small targets. An Efficient Multi-scale Attention (EMA) module is integrated to boost the model's comprehension of pixel-level interactions and overall contextual information. The implementation of the Dynamic Head (DyHead) enhances the representational capacity of the object detection head. These optimizations enhance the model's capability to capture complex features across different hierarchical levels and spatial locations. The model is trained on the S2TLD dataset, and the issue of class imbalance is addressed by data augmentation. The experimental findings indicate that the introduced algorithm markedly enhances the model's capacity to identify small traffic lights within complex environments, resulting in a 1.7% improvement in [email protected] over the baseline algorithm.